Loading…
Cortical surface registration network that learns which folds predict functional organization, using multi-modal supervision during training only.
No providers recorded yet. Browse all providers
Across most of the human cerebral cortex, a fold is a poor guide to where a cortical area sits. The central sulcus tracks primary motor and somatosensory cortex reliably, but elsewhere the relationship between folding and areal boundaries is inconsistent, and the folds themselves do not correspond one-to-one between individuals. Registration algorithms that align brains by folding alone therefore overfit: they sharpen sulci in the group average while smearing the functional networks researchers actually want to compare. The incumbent answer heavily regularizes folding-based registration, with that strength hand-tuned to maximize alignment of features the algorithm never sees.
DRUMS — Deep-learning Registration Using U-Net with Multimodal Supervision — moves the multi-modal information into the training objective rather than the inference input. The network reads folding features, sulcal depth and mean curvature, and is supervised by the cross-correlation of myelin maps, folding-corrected cortical thickness and 76 resting-state network components. It therefore learns which folds are trustworthy proxies for areal organization, and learns a spatially non-uniform regularization profile rather than requiring one to be tuned by trial and error. It was developed at Washington University in St. Louis by the group behind the Human Connectome Project, and is proposed as a replacement for Multi-modal Surface Matching (MSM), the classical per-pair optimizer that the HCP Pipelines currently use.
DRUMS operates on spherical cortical meshes, deforming vertices along the surface of a sphere. That makes it a sibling in kind, not a competitor, to the catalog's volumetric registration models — uniGradICON, BrainMorph and TotalRegistrator — which align 3D volumes.
The architecture has three stages — multiscale feature extraction, multiscale feature integration, and deformation-field generation — built around a five-level U-Net encoder spanning icosphere resolutions from ico-6 (40,962 vertices) down to ico-2. Convolution uses the Gaussian mixture model operator from MoNet, weighting each vertex against its neighbors in a mesh pseudo-coordinate space. Training used 1,071 WU-Minn HCP Young Adult subjects split into non-overlapping sets of 475 training, 475 testing and 50 validation, with family relationships kept from crossing the split; each subject contributed an 81-dimensional per-vertex feature vector. Optimization was AdamW with 15 warmup epochs and exponential moving averaging, for 500 epochs (DRUMS-Sulc) or 1,500 (DRUMS-SupSulc, DRUMS-All); DRUMS-All takes about two days on four NVIDIA A100 GPUs.
On held-out task-fMRI — 86 GLM maps never used in training — DRUMS-All at λ=0.01 reached a task cluster mass of 294K against MSMAll's 281K, with resting-state network cross-correlation of 0.697 versus 0.692. Among folding-input methods, DRUMS-SupSulc reached 0.677 RSN cross-correlation and 255K task cluster mass, against 0.641 and 248K for MSMSulc. On 45 test-retest subjects, individual areal classification agreed on 51.7k vertices across the 360 cortical areas versus 50.6k for MSMAll, with group overlap of 45.3k versus 43.3k.
The immediate consumer is the HCP Pipelines and the HCP-style processing they have seeded across large neuroimaging cohorts, where DRUMS-SupSulc is offered in place of MSMSulc and DRUMS-All in place of MSMAll. Better and more reproducible cortical alignment propagates into group fMRI analyses, individual cortical parcellation, myelin mapping and morphometry, and the second-per-subject cost makes it cheap to rerun across a whole cohort. DRUMS-Sulc has a narrower use: within-individual registration between surface meshes from different data types, such as low-field MRI or tissue photo reconstructions, where folding really does correspond.
DRUMS is the first learned cortical surface registration method the HCP group has advanced as a successor to its own MSM software, and it sets a baseline on HCP-style data under alignment, distortion and test-retest reproducibility together — the last of which the authors argue has been under-weighted in registration evaluation. Several caveats bound the result. The preprint has not been peer reviewed. Training, the template and every evaluation come from a single cohort of healthy young adults, and the authors state that DRUMS may not suit non-HCP-style datasets that fail its preprocessing prerequisites, such as fMRI denoising quality. No implementation code or trained weights accompany the preprint, so the reported numbers cannot yet be reproduced independently.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.